The Demand for Money in Tanzania · both serve as proxies for the transactions demand for money.2 The two variables behave similarly over time, but differences due to aid, terms of

www.theigc.org

Working Paper 11_0336 May 2011

International Growth CentreLondon School of Economics and Political Science 4th Floor, Tower Two Houghton Street London WC2A 2AE United Kingdom

For media or communications enquiries, please contact Adam Green [email protected]

Directed and Organised by

The Demand for Money in TanzaniaChristopher S. Adama, Pantaleo J. Kessyb, Johnson J. Nyellac and Stephen A. O’Connelld

aOxford and IGC (contact: [email protected]); bBank of Tanzania and IGC (contact: [email protected]); cBank of Tanzania (contact: [email protected]); dSwarthmore and IGC (contact: [email protected])

2

The Demand for Money in Tanzania

This paper is the outcome of research collaboration between staff of the Bank of Tanzania and the International Growth Centre. The views expressed in this paper are solely those of the authors and do not necessarily refl ect the offi cial views of the Bank of Tanzania or its management. All errors are those of the authors.

AbstractWe develop an econometric model of the demand for M2 in Tanzania, using quarterly data from 1998 to the present. The continuous decline in velocity since the late 1990s is associated with a transformation of economic activity that has cumulatively increased the monetary intensity of GDP. Portfolio behavior also responds to expected infl ation and to exchange rate depreciation, with weaker effects from interest rates. The components of M2 respond to opportunity costs as expected, with currency more sensitive to expected infl ation and deposits more sensitive to the interest rate on government securities. We discuss the policy implications of our results, including their relevance to the velocity-forecasting exercise that plays a key role in the Central Bank of Tanzania’s policy framework.

3


ContentsIntroduction 4

The Tanzanian setting X

Trends in velocity X

Composition of the aggregates X

Key features X

Existing work on money demand in Tanzania X

The demand for M2 X

Incorporating structural change X

Long-run estimates X

Short-run dynamics X

Currency and deposits X

Forecasting velocity X

Within-sample performance X

Out-of-sample forecasting X

Conclusions and Implications for Monetary Policy X

References X

Tables and Figures X

Appendix 1. Macroeconomic Developments, 1966-95 X

Appendix 2. Motivating the demand for money X

Appendix 3. Defi nitions and sources of variables X

Appendix 4. Phillips-Perron Unit root tests X


4

1. IntroductionThe Bank of Tanzania (BoT) uses a monetary aggregate as its intermediate target for monetary policy, on the grounds that portfolio equilibrium induces a reasonably predictable relationship between money and prices in Tanzania. The fulcrum of this relationship is the private sector’s demand for money. If this is a stable function of observable variables, then a policy that targets the growth of nominal money has some prospect of stabilizing infl ation at desired levels and at reasonable cost in terms of other variables. But if the demand for money is subject to large and unpredictable shifts, an approach that places less emphasis on money growth may produce superior macroeconomic outcomes. Instability of money demand is widely viewed as having contributed to the demise of money-targeting frameworks among the industrial and emerging-market economies and their replacement since the early 1990s with variants of infl ation targeting (Freedman and Laxton 2009).

While the demand for money has been the subject of considerable research within Tanzania, data constraints are severe and the published literature is relatively small. Econometric models currently do not play as prominent a role in policy formation as the Bank desires. Forecasts of nominal money demand are required to determine program targets for money base growth, but these generally come down to judgmental extrapolations of trends in velocity.

We pose a simple question in this paper: does a stable money demand function exist in Tanzania? Given its prominence in Tanzania’s monetary framework we focus primarily on broad money (M2). Currency is an unusually large proportion of M2 in Tanzania, however, and recent data suggest that fewer than 10 percent of rural households have a member with a bank account. We therefore disaggregate M2 and present results separately for currency and deposits. Further work is a high priority, including an investigation of M3, which includes foreign-currency deposits, the fastest- growing component of bank liabilities since the mid-1990s.

We begin the paper by summarizing some of the key features of portfolio behavior in Tanzania. Section 3 reviews existing econometric work on money demand. In Sections 4 and 5 we present our empirical work, focusing in turn on long-run relationships and short-run dynamics. Section 6 discusses the relevance of our results to the velocity forecasts that underpin the BoT’s reserve-money program. We conclude in Section 7 with a summary of fi ndings and policy implications.

5


2. The Tanzanian settingDenoting the log of real M2 by a conventional demand for money function takes the form

(1)

where is a scale variable, typically the log of real GDP, and are row vectors of interest rates on alternative assets and the components of M2, E[.] is the expected value operator, is the infl ation rate, is the rate of nominal depreciation, and z is a vector of other determinants (Ericsson 1998).1

As alternative scale variables, we consider real GDP and gross national expenditure (GNE); both serve as proxies for the transactions demand for money.2 The two variables behave similarly over time, but differences due to aid, terms of trade effects, and private capital fl ows can sometimes be substantial. In the end our empirical results favor real GDP.

The menu of alternative assets includes claims denominated in foreign currency, domestic government securities, and inventories of goods. Foreign currency was widely traded during the exchange-control period and has been legal since the early 1990s. Foreign-currency deposits were introduced in the early 1990s and now account for nearly a third of total M3. Capital controls continue to prohibit the accumulation of offshore assets, but the effectiveness of these controls is uncertain and there may well be substantial dollar balances abroad. We use expected depreciation as our main measure of the return on assets denominated in foreign currency. Domestic government securities are limited to short-maturity Treasury bills, introduced in 1994. To capture the return on inventories of goods, we use the expected infl ation rate, a variable that features prominently in money demand work on developing countries (Siram 2001).

Our econometric work focuses primarily on the period since 1998. While the brevity of the sample is a serious limitation, there is a tradeoff between the benefi ts of longer runs of data and the misspecifi cation that can arise in fi tting empirical models across very different economic regimes (Juselius 2006). As outlined in Appendix I, the decade that preceded 1998 was one of intensive economic reforms. While many of these were in place by 1994 – including interest rate liberalization, exchange rate unifi cation, the removal of price controls, the licensing of foreign exchange bureaus, the introduction of foreign currency deposits and treasury-bills, and the opening of the banking sector to competition – the transformation of the public sector took longer, and occupied much of the 1990s. The near-monopoly National Bank of Commerce (NBC), for example, was restructured repeatedly during the fi rst half of the 1990s, and in 1997 was relieved of most of its rural branch network in a split that created the new National Microfi nance Bank. The new NBC was fi nally privatized only in 2000, and the NMB in 2005 (Cull and Spreng 2010). Other public enterprises continued to be privatized throughout the 1990s, with the fi rst 115 divested between 1992 and 1994 and another 168 between 1995 and 1998 (PSRC 2000; Waigama 2008).

Trends in velocity

Figure 1 shows the ratio of real GDP to real money balances – the velocity of money3 – for the four main monetary aggregates in Tanzania, from the late 1980s to the present. The sharp increase in 1995 coincides with the adoption of a cash budget by the public sector, a move Ndulu (1997) credits with validating a disinfl ation strategy that was based on restricting growth in the monetary aggregates (at that time M3 was the main target). However, while the initial increase in velocity is a plausible effect of tight money, its persistence through most of the

1 The theory underlying this specifi cation is discussed in more detail in Appendix 22 Theory suggests a role for fi nancial wealth as well, and recent work on South Africa confi rms its potential importance (Hall et al 2009),

but wealth data are unavailable.3 Velocity is usually measured as the ratio of nominal GDP to the nominal money stock. Here we measure it as the ratio of real GDP to

the real money stock. This is necessitated by our choice to interpolate real GDP from annual to quarterly frequency: we do not have a reliable quarterly nominal GDP series. The two measures differ by the ratio of the GDP defl ator to the consumer price index.

6


second half of the 1990s is puzzling. None of the standard determinants of money demand deteriorated during this period: real GDP was rising (in contrast to the more typical contraction in a money-based stabilization), infl ation was falling, nominal depreciation remained modest, and interest rates fell rapidly from their initially high levels.

Problems in the banking sector may have played some role in dampening money demand during the mid to late 1990s. The restructuring process, in particular, appears to have restricted the access of households to fi nancial services, particularly in rural areas (Cihak and Podpiera 2008). In 1994 alone, the NBC retrenched 2,800 employees and closed 23 branches. Successive household budget surveys show a sharp decrease during the 1990s in the proportion of households reporting a bank account – this number falls from 18.0 percent in the 1991/92 survey to 6.4 percent in 2000/01, before recovering to 10.0 percent in the 2007 survey.4

Confi dence in the banking system may also have been shaken by the collapse of Meridien-BIAO and Tanzania Housing Bank in the mid-1990s, though these failures do not appear to have systemically endangered the banking system.

The trend of fi nancial shallowing reverses itself around 1999/2000, and fi nancial deepening continues strongly throughout the ensuing decade. The ratio of M2 to GDP exceeds 20% for the fi rst time in 2008 – very low by international standards, but a typical value for low-income Africa (Honohan and Beck 2007). Below we show that a variety of measures of structural change reverse course during the second half of the 1990s, consistent with a fundamental reorientation of the economy towards trade, investment, and private sector activity.

Composition of the aggregates

M3 is defi ned in Tanzania as M2 plus foreign currency deposits. The composition of M3 (Figure 2) brings out two main observations. The fi rst is the continued prominence of currency – at 20 percent of M3 in 2008 – despite its gradual replacement by bank deposits. Currency remains the dominant means of exchange in much of Tanzania, especially in rural areas.5 Figure 3 provides some evidence of the links between currency and rural economic activity in Tanzania. The seasonal pattern of real currency holdings is very strongly correlated with the seasonal in the monthly real price of food, where the latter is defi ned as the food CPI defl ated by the non-food CPI. Nyella (2005) suggests that these co-movements are driven by the crop cycle, with currency holdings peaking at the time of harvest (he also cites an end-of-year holiday spending effect).

The second feature of M3 is the growing importance of foreign-currency deposits (FCDs), which peaked in 2006 at just above one third of the total. Kessy (2008) fi nds that the share of FCDs in total deposits is an increasing function of expected depreciation and the share of trade in GDP. Figure 4 compares the ex post yield on 12-month foreign currency deposits with the interest rate on comparable domestic deposits over the past decade. For most of the period FCD have constituted a favorable medium-term store of value for households prepared to tolerate the volatility in their return.

4 It is likely that the proportion bottomed out in the mid- to late-1990s, as NBC was being restructured and new banks had not yet established a substantial foothold. Reductions in rural access may well have lasted longer, however. The National Microfi nance Bank, which succeeded NBC and acquired most of NBC’s rural branch network, remained in state hands until 2005, operating for most of the 1997-2005 period under a private consulting contract that focused on enhancing profi tability, not on expansion of reach or services. See Cull and Spreng (2008).

5 While 10 percent of Tanzanian households reported a member with a bank account in 2007, this proportion was nearly 25 percent in Dar es Salaam. The fi gure for rural areas must therefore be well below 10 percent.

7


Key features

Our discussion has emphasized a set of potentially important features of the Tanzanian case since the mid-1990s. These include

• a large share of currency in the higher monetary aggregates;

• a small share of the public with bank accounts, particularly in rural areas where the economy remains largely cash-based;

• the coexistence of deposit dollarization with capital controls, potentially allowing a higher degree of exchange-rate-based portfolio substitution within M3 than between domestic and foreign bonds;

• the introduction of Treasury Bills in 1994, offering a high-yielding alternative to bank deposits;

• signifi cant structural change in the economy stemming from macroeconomic and supply-side reforms which shifted economic activity towards more monetary-intensive activities and;

• the transformation of the banking sector after 1992, which improved the quality of bank services while potentially restricting access to banking services at least initially.

The reforms of 1988-94 changed the menu of fi nancial assets and the character of the banking sector. In our view, these reforms were fundamental enough that an investigation of contemporary portfolio behavior should start in 1994 at the earliest. We also noted, however, that velocity behaves very differently during the second half of the 1990s than subsequently. In principle, the sharp change in money supply behavior starting in 1995 should assist in the econometric identifi cation of the money demand equation. As we have stressed, however, the disinfl ation period as a whole remains something of a puzzle. Real money balances fell sharply and persistently, despite the absence of commensurate movements in the conventional determinants of demand. Below we identify components of the vector that perform strongly in the post-1998 sample while providing at least a partial account of the collapse in real money demand during the second half of the 1990s.

3. Existing work on money demand in TanzaniaWe briefl y review the Tanzanian literature, beginning with work that preceded the fi rst application of modern time series methods by Randa (1999).

The early literature features a set of partial adjustment models applied to annual data (Gerdes 1990, Maje 1992, Kihuale 1994, Mgonya 1997). Estimation samples start in 1967 and end in 1995 or earlier. These models incorporate a lagged dependent variable but otherwise have very sparse dynamics, if any. They use expected infl ation as the opportunity cost of money, refl ecting the undeveloped nature of the fi nancial system and the absence of interest-bearing alternatives to money. In the earliest study, Gerdes (1990) introduces the key empirical themes of this literature: high income elasticities, even by the standards of low-income countries; conventional but weak portfolio responses to expected infl ation; and controversies over the empirical stability of the estimated relationships.

These early contributions uniformly employed real GDP as the scale variable. Using data from 1967 to 1985, Gerdes (1990) reported long-run income elasticities of close to 2.0 for M1 and M2, against a benchmark of 0.5 from the Baumol-Tobin transactions demand model and 1.0 from the quantity theory of money. In line with the literature on low-income countries, he attributed these high elasticities to ongoing monetization of the economy and the persistence of limitations on the menu of alternative fi nancial assets. Similar long run income elasticities were found in subsequent applications, including prominently that of Maje (1992), who incorporated the number of branches of the monopoly National Bank of Commerce (NBC) as a direct proxy for the pace of monetization. The sign on this variable was positive, as expected, but was not statistically signifi cant.

8


The coeffi cients on expected infl ation varied across applications, but they tended to be small and were in most cases statistically insignifi cant. Starting with Maje (1992) the time deposit rate was often included as a measure of the return to alternative assets (M1) or as part of the own return on money (M2); point estimates were often appropriately signed, but were uniformly small and imprecisely estimated. It was not possible in this early literature to reject the hypothesis of a perfectly inelastic response of money demand to interest rates.

Gerdes (1992), Maje (1992) and Mgonya (1997) all implemented Chow tests to assess the stability of the money demand function, as did Kihuale (1994) who applied a broader battery of tests using data from 1967 to 1990. Results were mixed; Gerdes and Mgonya were unable to reject parameter stability, while Maje and Kihaule found evidence of instability. While the latter fi nding was perhaps more plausible in a highly controlled economy undergoing the stresses of fi nancial repression, economic collapse, and structural adjustment, the sources of instability were not carefully diagnosed. A limitation of this early literature, moreover, is the diffi culty of distinguishing genuine parameter instability from slow equilibrium correction when the variables have stochastic trends. More generally, conventional tests of signifi cance can be misleading when the variables are non-stationary, as emphasized in the time-series literature following Engle and Granger (1987); a cointegration/error-correction framework is required to correctly identify short- and long-run relationships and support accurate inference (Enders 2007).

Randa (1999) established the standard for contemporary research by moving to quarterly data (1974q1-1996q4), incorporating a role for currency substitution, and applying modern time series methods. Randa focused on M0, M1 and M2; Nyella (1998) applied similar methods to M3, using quarterly data from 1986q1 to 1997q4. The hallmark of the modern time-series approach is two-fold: fi rst, close attention to the stationarity properties of the data and the resulting distinction between long-run equilibrium relationships and short-run dynamics; and second, where feasible, the adoption of a system-based approach in which multiple long-run relationships may exist among the variables. We return to these themes in Section 4 below.

Moving to quarterly data requires interpolation of the scale variable, since the GDP accounts are available only annually. Randa uses quadratic interpolation, which does not require (or benefi t from) the use of indicator variables available at quarterly frequency. Using similar interpolation methods, Nyella uses gross national disposable income (GNDI), which differs from GDP by the sum of net factor income and net transfers, a quantity dominated by large and variable amounts of foreign aid.

Randa’s second innovation was the incorporation of the rate of depreciation of the black market exchange rate as a measure of opportunity cost for households with access to assets denominated in foreign currency. Currency substitution was illegal until the early 1990s, and it is still illegal for Tanzanian residents to hold offshore assets. The foreign exchange black market was nonetheless active starting in the mid-1970s, and there is considerable evidence that portfolio infl uenced the behavior of the parallel premium during the 1970s and 80s (O’Connell 1992).

Randa’s specifi cation omitted interest rates in favor of infl ation and depreciation. The Johansen method confi rmed the presence of a single cointegration vector relating real money balances, real GDP, infl ation and depreciation. Interpreted as a money demand function, all coeffi cients had the expected signs and were statistically signifi cant. Estimated income elasticities were higher for successively broader monetary aggregates, a fi nding also characteristic of the earlier literature. In all cases, however, they were markedly smaller than previous work had suggested (0.81, 0.96 and 1.11 for M0 M1 and M2 respectively) and more in line with estimates for other countries. The infl ation elasticity of demand for money was negative, as expected (magnitudes 0.43, 1.08 and 1.16 for M0, M1 and M2, respectively), as was the elasticity with respect to expected depreciation.

Randa interprets stability in terms of existence of a cointegrated long-run money demand relationship. His conclusion is that a stable set of demand functions exists despite the economic and fi nancial reforms that have taken place since late 1980s. Nyella (1998) reports a similar fi nding

9


for real M3 (M2 plus foreign currency deposits), using quarterly data for 1986q1 to 1997q4. Nyella’s specifi cation incorporates a measure of the own return on money, in the form of the time deposit rate adjusted by the ratio of these deposits to M3. It also incorporates expected depreciation, which plays double duty as a component of the own return on foreign currency deposits and an opportunity cost for households with access to foreign currency and offshore deposits that are outside of M3.

Nyella (1998) innovates by incorporating a deterministic proxy for fi nancial liberalizations undertaken in the early 1990s. This proxy goes from zero to 1 when foreign currency deposits are introduced in 1992, and from 1 to 2 when the government T-bill market is introduced in 1994. This variable is constrained to lie in the cointegration space, implying a sequence of two equal and permanent shifts in the long-run relationship.

As in Randa (1999), Nyella’s analysis supports the presence of a cointegration vector with the characteristics of a long-run money demand function. At 0.58, however, the income elasticity is extremely low by the standards of previous research. The own rate of return has a positive and statistically signifi cant impact as expected, and the coeffi cient is large, at −3.7. Depreciation has a positive (0.9) and signifi cant coeffi cient, consistent with stronger substitution into domestic foreign currency deposits than into other foreign assets when expected depreciation rises. Short run elasticities from the error-correction representation of the money demand equation were plausibly signed in most cases.

4. The demand for M2We estimate the demand for the M2 aggregate and its two sub-components, currency and deposits, using quarterly data on the variables identifi ed in equation (1) above (see Appendix 3 for details). Quarterly data are available for all monetary aggregates, prices, interest rates and exchange rates. Data on the scale variables (real GDP and real GNE) are only available at an annual frequency. We interpolate these to a quarterly frequency using an augmented version of the ‘proportional Denton interpolation’ method.6

The modern approach to money demand starts by interpreting equation (1) as a long-run equilibrium relationship. The deviation from this equilibrium should therefore be a stationary random variable with a zero mean. Denoting the vector of determinants by

and assuming

linearity, this implies

(2)

where the ‘equilibrium error’ has a zero unconditional mean and constant unconditional variance and where the sub-vector may include deterministic components (e.g., a constant, a linear trend, seasonal dummy variables). We verify below that m and the variables in w have unit roots (see Appendix 4). In this case the stationarity of u implies that the variables are cointegrated. The long-run money demand coeffi cients correspond to the cointegration vector

If there is only one long-run relationship between the variables in then the long-run coeffi cients can be estimated consistently in a variety of ways, including OLS as applied to equation (2). The short-run dynamics of the system can then be recovered by estimating an error-correction specifi cation in which the change in money balances responds to current and lagged changes in the other variables as well as to the previous period’s estimated equilibrium error,

. (3)

6 This method generates a smoothed interpolated quarterly series from annual data subject to the constraint that the sum over the year of the interpolated quarterly values for GDP equals the known annual total GDP. We augment this series by exploiting known quarterly movement in correlates of real GDP, such as trade fl ows and investment expenditure.

10


Cointegration implies that the adjustment speed lies between −2 and 0. While OLS yields a consistent estimator of the adjustment speed, consistency of the other estimates relies on weak exogteneity of the right-hand-side variables (Enders 2006).

In large samples, a systems approach is likely to be preferable to the Engle-Granger two-step method we have just described. Starting with an unstructured vector autoregression involving the vector the Johansen method allows joint estimation of the short- and long-run coeffi cients and provides an orderly approach to assessing exogeneity and identifying multiple cointegrating relationships among the variables (e.g., Juselius 2006). In the presence of a single cointegration vector, however, the Engle-Granger two-step method tends to produce smaller bias in very small samples. In what follows we rely on the two-step procedure, leaving systems estimation as a potential extension.

Incorporating structural change

As emphasized above, the money demand literature in Tanzania has tended to feature income elasticities that are both very high and sensitive to the sample. The standard explanation for high income elasticities in low-income countries is that increases in aggregate income are accompanied by structural changes that increase the use of currency and bank deposits in transactions. These changes can bias upwards the measured income elasticity by confl ating it with the impact of structural change.

While the ‘monetary intensity’ of economic activity cannot be measured directly, we can proxy it based on the set of measures depicted in Figure 5. Starting from the activity side, we defi ne the monetary-intensive share of GDP as the share of mining and quarrying, manufacturing, electricity and water supply, trade, restaurants and hotels, transport and communications, and fi nance, insurance and real estate. We then include investment as a share of GDP and imports as a share of GDP, on the grounds that these components of expenditure are more transactions-intensive than other components of GDP. The government wage bill as a share of GDP is next; this has increased substantially in recent years, refl ecting increases in both real wages and employment (including MDG-related spending). In contrast to much of the private sector, government wages and salaries are paid though the banking system. Finally, credit to the private sector as a share of M3 provides a direct measure of banking sector intermediation.

These proxies tend to emphasize change in the demand for liquidity services. Other things equal, reductions in the effective cost of acquiring fi nancial services would also serve to increase the monetary intensity of economic activity, and vice versa. To date, however, it has not been possible to compute measures of these supply-side changes, either directly in terms of the cost of access to fi nancial services or indirectly through measures of the changing structure and operations of the fi nancial system as a result of the reforms described in Appendix 1. We hope to develop such measures and incorporate these in our measure of structural change in subsequent work.

While each of these variables captures a distinct aspect of the changing monetary intensity of GDP, data limitations preclude our incorporating multiple proxies in our regressions. We can summarize the information content in these measures, however, by extracting their fi rst principal component. Table 1 shows the results of this exercise. There is a very strong commonality across the movements of these variables, such that a more-or-less equally weighted average accounts for nearly three-quarters of their total covariance. We use this weighed average as our proxy for transactions intensity. As indicated in Figure 5, transactions intensity falls dramatically in the early stages of the 1995-99 disinfl ation, before stabilizing in the late 1990s and then rising sharply and cumulatively starting around 2002.

11


Long-run estimates

Table 2 shows our results for the long-run demand for M2, for samples beginning in 1994q1 and 1998q1. Note that in all specifi cations a Dickey-Fuller test rejects the hypothesis of a unit root in the residuals. Since the variables themselves are I(1), this is evidence of cointegration. Results are somewhat stronger for real GDP than for real GNE, and in subsequent tables we focus solely on specifi cations involving real GDP.

The results for the post-1998 sample appear in columns 3 and 4. The transactions intensity of GDP comes in very strongly, and with this dimension of structural change included in the regression, the expenditure and income elasticities – at 1.58 and 1.78 when transactions intensity is excluded – fall sharply. The scale elasticities are now insignifi cantly different from 1. Expected infl ation and expected depreciation have the anticipated signs and reasonable magnitudes, with a 10 percent increase at an annual rate reducing real money demand by 0.8 and 2.2 percent, respectively. 7

We noted earlier that an existing literature fi nds weak interest rate effects, based mainly on samples ending in the mid-1990s or earlier. Perhaps surprisingly, this effect persists after a decade of banking sector reforms. The spread between the T-bill rate and the rate on time deposits has a plausible sign and magnitude in column 3, but neither survives in column 4, and we cannot reject that the coeffi cient is zero in either case. At least one of the centred seasonal dummies, in contrast, has a large and statistically signifi cant coeffi cient in each specifi cation.

Figure 6 shows the cointegration relationship for the regression in column 4. While real money balances and their determinants move closely together, the mean squared error is substantial at 3.7 percentage points. Equilibrium errors are sizeable over the 2005-2007 period, with a persistent under-prediction starting in 2005 followed by over-prediction in 2007.

As shown in columns 1 and 2, cointegration survives but the fi t deteriorates markedly when the sample is extended back to 1994. The root mean squared errors rise by roughly a third. Among the opportunity cost variables, only expected depreciation continues to perform strongly, its signifi cance intact and its magnitude roughly doubling. The income elasticities are now signifi cantly below unity.

Short-run dynamics

In Table 3, we estimate dynamic error-correction models (ECMs) using the estimated equilibrium errors from columns 2 and 4 of Table 2. Note that all variables in these specifi cations are stationary – the equilibrium errors by virtue of cointegration, and the remaining variables via differencing. Statistical tests suggest excluding the transactions intensity variable from the short-run dynamics, and so we proceed accordingly. We focus on the post-1998 sample (columns 2 and 3).

Column 2 shows an unrestricted ECM that includes one lag of all variables, and in column 3 we reduce this to a parsimonious specifi cation with stronger statistical properties.8 The parsimonious specifi cation generates a root mean squared error of 2.6 percentage points, by comparison with a standard deviation of the dependent variable of 3.6 percentage points per quarter. There is no evidence of serial correlation in the post-1998 sample.

7 For all models we report heteroscedasticity-robust standard errors. Our interpolation of real GDP allows us to exploit a sample of approximately 40 observations. However, this masks the fact that we have only 11 independent observations on real GDP. To check that this does not seriously distort our interpolation, we also compute standard errors from a bootstrap routine. This slightly raises the estimated standard errors reported in the tables in this paper but not suffi ciently to alter any inference we draw from the results.

8 Where appropriate we accommodate higher lags in the parsimonious specifi cation

12


Error correction is an implication of cointegration, and so the negative and statistically signifi cant coeffi cient on the lagged equilibrium error is important corroboration of our cointegration result. The speed of error correction is considerable, at over 30 percent per quarter. At the same time the growth in real money balances displays considerable short-run inertia, with a lag coeffi cient of roughly 0.3. Real GDP has a smaller impact in the short run than it does over time, but the coeffi cient is strongly signifi cant. Infl ation and depreciation both come in strongly in the short run; and the interest-rate spread continues to be insignifi cantly different from zero. The seasonal dummies remain large and important.

Figure 7 shows the actual and fi tted values from the parsimonious ECM (column 3). The short-run dynamics go some way towards resolving the large and persistent equilibrium errors in 2005-2007, with the exception of a single large outlier in the fi rst quarter of 2005.

5. Currency and depositsIn Tables 4-6 we undertake a similar exercise for the two main components of M2, currency and domestic-currency bank deposits. Our aim here is to shed some light on compositional effects within M2. While these effects are present in any money demand exercise, they may be of particular interest in Tanzania. The dominance of currency in rural areas, the small proportion of the population with bank accounts, and the potentially even smaller proportion with interest in or awareness of the T-bill market, suggests that the relevant opportunity costs may differ sharply across the components of M2. The puzzle of largely-absent interest-rate effects, in particular, may in part be a result of aggregation, if markets are partly segmented and only a portion of M2 responds to interest rate differentials.

While the results should be considered preliminary, they suggest sharp and intuitively appealing differences within M2. As earlier we focus on the post-1998 results. In Table 4, long-run currency demand responds strongly to expected infl ation and not at all to interest rates (we exclude them on statistical grounds). Perhaps surprisingly, expected depreciation also has a small and insignifi cant impact on currency demand. The demand for deposits, in contrast, responds very strongly to expected depreciation, consistent with substitution between domestic- and foreign-currency deposits within M3. The point estimate on the interest rate spread (we use the T-bill rate minus a weighted average of domestic-currency deposit rates), moreover, is considerably larger in the deposit equation than in the overall M2 regression, and much closer to statistical signifi cance. The impact of expected infl ation on deposits is small and statistically insignifi cant. A weak response to interest rates remains a feature of the short-run dynamics, however, and infl ation appears to have at least as strong a short-run impact on domestic-currency deposits as it does on currency (Tables 5 and 6)

6. Forecasting velocityForecasts of velocity play a central role in the Bank of Tanzania’s policy framework. Expressed in terms of growth rates, the defi nition of velocity ( implies that

(4)

where v is the log of velocity. The change in real money balances, in turn, is the difference between nominal money growth and infl ation: The Bank of Tanzania’s reserve-money program combines an infl ation target with a projection for real growth to yield an infl ation-consistent growth rate for nominal money:

(5)

13


The velocity forecast therefore directly affects the programmed growth rate for M2.

Large forecast errors are potentially costly. Too high a projection for implies a tighter policy stance than intended: failure to accommodate a higher-than-anticipated money demand may drive infl ation below target, while serving as a brake on real economic activity. Too loose a projection may feed infl ation or asset price bubbles, with delayed adverse effects on the economy.

The Bank of Tanzania’s current approach relies on extrapolation of recent trends in velocity. This approach makes sense when velocity appears to be subject to relatively slow-moving trends: as indicated in Figure 8, velocity has been falling at an average of 5.25 percent per year for almost a decade. There is substantial short-run volatility around this average, however – the standard deviation of the year-on-year change is 6.1 percent.

In this section we briefl y compare the Bank’s trend-extrapolation methods with alternative approaches based on the time-series behavior of velocity, including a VAR incorporating the determinants of money demand. We end by discussing how information from an econometric money demand equation can be used in a velocity-forecasting exercise. We proceed in two steps, fi rst looking at the within-sample performance of alternative models and then considering their out-of sample forecasting properties. The fi rst of these steps is diagnostic: it allows us to assess and discuss the properties of alternative models against the background of full information about the data.

Within-sample performance

The BoT’s approach is an example of a broader class of univariate forecasts that rely only on the past behavior of velocity. In what follows we compare a set of univariate approaches with a multivariate approach based on the variables in our money demand equation. We begin with a pair of univariate models that assume a locally deterministic path for velocity. With some oversimplifi cation, these represent alternative formalizations of what the BoT does in practice.

(i) A rolling trend estimator assumes that velocity will return, next quarter, to the seasonally-adjusted trend line it has followed over the recent past. For purposes of comparison, we use a 3-year window. The rolling trend estimator for period t is therefore the one-period-ahead extrapolation of a linear trend estimated over quarters t−13 to t−1, with seasonal dummy variables included.

(ii) A moving average growth estimator extrapolates the growth rate of velocity, rather than its level. A velocity forecast of this type features prominently in the ‘McCallum rule’ for base money growth (McCallum 1988). Maintaining our 3-year window, we compute the moving average growth estimator for quarter t as the value of velocity in quarter t−4, plus the average year-on-year growth in velocity over quarters t−13 to t−1.

We also consider a pair of univariate models that treat the trend in velocity as stochastic.

(iii) A simple random walk with drift (and deterministic seasonal effects) takes the form

(6)

where the parameters are estimated over the full sample. This model has the well-known property that the best estimate of velocity next period is its value this period (after accounting for the estimated drift and seasonal determinants of velocity).

14


(iv) A random walk with drift is a special case of a more general set of I(1) models of which the stationary part may exhibit a combination of autoregressive and/or moving average components. The Box-Jenkins method provides a set of guidelines for identifying a parsimonious ARIMA model that ‘best’ exploits the observed autocorrelations and partial autocorrelations of a given time series. Estimating over the period since 1998q1, we selected an ARIMA(1,1,2) model with deterministic seasonal factors. This model takes the form:

.

(7)

Univariate forecasting models employ only the history of velocity itself. Our money demand equation, however, suggests a natural multivariate alternative. Using r to denote the non-income determinants of the demand for money, equation (3) can be written as

, (8)

where for simplicity we have imposed a lag order of p = 1. Using the defi nition equation (8) implies the velocity equation

(9)

Equation (9) can in turn be rewritten as a distributed lag model in which the level of velocity depends on its own lags and on the current and lagged levels of the determinants of money demand In effect, we can think of (9) as the fi rst row in the structural simultaneous equations system

(10)

where the vector includes velocity and the determinants of money demand discussed above, and is a vector of structural disturbances.

Equation (10) is a structural simultaneous equations model, but as long as B is invertible it will have a reduced-form VAR representation of the form

(11)

where and . Equation (11) is stochastically balanced, and therefore can be estimated consistently in levels, as long as each I(1) variable in the VAR is cointegrated with at least one other I(1) variable in the VAR – a condition our cointegration analysis has established. The estimated VAR provides a basis for multivariate forecasting of velocity – or, for that matter, of any element of the vector xt. It constitutes our fi nal forecasting model.

15


Table 7 and Figures 9a and 9b compare the within-sample properties of the four univariate approaches and the VAR.9 We report two standard measures, the root mean square error (RMSE) and the mean absolute error (MAE).10 Amongst the univariate models, the time-series models systematically out-perform those that treat the trend as locally deterministic. All models will fail to predict turning points in the data, but short-lag time-series based models will ‘get back on track’ more rapidly than deterministic-trend models (this difference will increase the longer the time-span used to estimate the local trend).

Figure 9b, which compares the goodness of fi t of the preferred ARIMA model – the best-fi tting univariate model of velocity – with that of the VAR model, clearly shows the substantial gains to conditioning the estimate of velocity on the full set of determinants of money demand. Prediction errors are generally smaller and shorter-lived. This increase in predictive accuracy is direct refl ected in Table 7 where the root mean square error of the VAR forecast is less than half that of the rolling trend (from 5.3percent per quarter to 2.4 percent per quarter) and more than one percentage point per quarter lower than the ‘best fi tting’ univariate ARIMA model.

Out-of-sample forecasting

Next we compare the out-of-sample forecast performance of the same set of models. We estimate each model over the sample period to 2006q4 and compute summary measures of forecast performance over the period from 2007q1 to 2008q4, reporting both one-step-ahead and, for the three time series models, dynamic 8-step-ahead forecasts for the period 2007q1 to 2008q4.11

Table 8 reports the forecast statistics which serve to reinforce the within-sample evidence. In terms of one-step-ahead forecast performance, the time-series models again dominate the deterministic-trend models while the velocity forecast from the VAR model again exhibits a substantially lower forecast error than any of the univariate measures. Conditioning on the determinants of money demand clearly and decisively enhances forecast performance. This is so for both the one-step and multistep forecast even though the latter are necessarily higher. Figure 10 illustrates the one-step-ahead forecast error for three representative measures (the rolling trend, the ARIMA and the VAR model). All three models initially under-predict velocity in the fi rst two quarters of 2007 and over-predict throughout the remainder of 2007 and into 2008, although the VAR again shows lower deviation and more rapid error correction. Finally, Figures 11 plots the multi-step forecast for velocity, estimated from the VAR, against the actual path of velocity, illustrating the systematic over-prediction of velocity throughout late 2007 and 2008.

7. Conclusions and Implications for Monetary PolicyWe have identifi ed a well-behaved dynamic demand function for M2 in Tanzania for the period from 1998 to the present and have shown how this model can be used to enhance the forecasting of velocity. Real income growth and structural change continue to generate a strong underlying trend of fi nancial deepening, and by separating these two effects our results shed light on the high and variable income elasticities that have been a feature of the Tanzanian literature. Conventional opportunity cost effects are present, but as in the earlier literature they appear to be dominated by substitution between domestic-currency and foreign–currency assets, and between money and infl ation hedges. Disaggregating currency and deposits, we fi nd that currency responds more strongly to expected infl ation, and deposits to the interest-rate spread vis-à-vis T-bills, than does overall M2.

9 By construction, the rolling trend and McCallum rule forecasts differ from the time series model in one important sense. By construction, for the former the forecast at time t is based only on information up to time t-1. This contrasts with the time-series models where the predictive performance of the model at time t is conditional on data up to time t-1 but the parameters of the models are estimated on data for the whole sample t= 1...T. This distinction disappears when we compare out-of-sample forecast performance.

10 The RMSE is the most commonly used measure, having the property that it is measured in the same units as the dependent variable itself. The MAE is similarly measured. The RMSE is, however, more sensitive than the MAE to outlier errors: the squaring process gives disproportionate weight to very large errors. Hence if the cost of an error is roughly proportional to the size of the error, not the square of the error, then the MAE may be a more relevant criterion. Generally, however, these statistics will vary in unison.

11 A one-step-ahead forecast is generated by estimating the model on data up to period t and forecasting the outcome for t+1 The estimate for t+2 is obtained by using the same model but computing the forecast conditional on the actual data up to t+1. By contrast, an h-period-ahead forecast computes each successive forecast period from t+1 to t+8 without updating the data beyond that available at time t. Thus, in this case, the estimate for t+2 is obtained using the data up to period t and the forecast made at time t of the value of the data at t+1. Forecast errors are therefore cumulated as the horizon is extended.

16


While our results go some way towards resolving the puzzle of persistently declining velocity during the disinfl ation of 1995-99, our empirical models deteriorate markedly when the four years 1994q1-1997q4 are included in the sample. Further work may well improve our account of this period, but for operational purposes our view is that the BoT should base its statistical work on the later period.

A variety of extensions are of high priority. The fi rst is to complement – or replace – our single-equation analysis with a multivariate approach that allows an orderly treatment of exogeneity issues and multiple long-run relationships. It remains to be seen whether our short sample will support such an approach, but some of the most important policy questions have to do with system-wide dynamics. For example, we have shown clearly that expected infl ation affects money demand; for policy purposes, the urgent question is whether monetary disequilibrium affects future infl ation. An integrated treatment of money and infl ation dynamics would be of considerable value.

A second extension is to apply the methods of this paper to M3. Kessy (2008) fi nds considerable evidence of substitution within M3, and our results are consistent with this. Understanding the behavior of overall M3 may be of considerable value, particularly in advance of capital account is liberalization.

Velocity forecasts play a central role in the BoT’s monetary policy framework, as they do throughout much of Sub-Saharan Africa. We have shown, not surprisingly, that a vector auto-regression model, based on our structural money demand equation, substantially out-performs a variety of univariate approaches, both within-sample and over a short out-of-sample horizon. Thus, for short-term forecasting purposes, the existence of a stable cointegrating relationship between real money balances – or velocity – and the determinants of money demand suggests that VAR-based forecasts, based on the most up-to-date data, may have substantial value in program formulation, as a complement to judgment and a check on simple univariate extrapolation.

17


ReferencesBevan, David, Paul Collier and Jan Willem Gunning (1993) Controlled Open Economies: A Neoclassical Approach to Structuralism (Oxford: Clarendon Press)

C ihák, Martin and Richard Podpiera 2005 “Bank Behavior in Developing Countries: Evidence from East Africa” IMF Working Paper 05/129 Washington, DC: International Monetary Fund, June

Cull, Robert and Connor Spreng 2008 “Pursuing Effi ciency While Maintaining Outreach: Bank Privatization in Tanzania” World Bank Policy Research Working Paper 4804 Washington, DC: The World Bank

Collier, Paul and Jan Willem Gunning (1991), ”Money Creation and Financial Liberalization in A Socialist Banking System: Tanzania 1983-88” , World Development. 19(5), 533-538

Enders, Walter (2007) Applied Econometric Time Series 2nd Edition, Wiley, New York

Ericsson, Neil R. (1998), “Empirical Modeling of Money Demand” Empirical Economics 23: 295-315

Freedman, Charles and Douglas Laxton (2009), Why Infl ation Targeting? IMF Working Paper 09/86, April.

Gerdes, William (1990), “The Demand for Money in Socialist Tanzania” Atlantic Economic Journal 18(3), September

Hall, Stephen, G. Hondroyiannis, P. Swamy, and G. Tavlas (2009) “Where Has All the Money Gone? Wealth and the Demand for Money in South Africa” Journal of African Economies 18: 84 – 112

Honohan, Patrick and Torsten Beck (2007) Making Finance Work for Africa World Bank, Washington, DC

Juselius, Katarina (2006) The Cointegrated VAR Model: Methodology and Applications Oxford: Oxford University Press

Kessy, Pantaleo (2008) “Empirical Evidence on the Determinants of Dollarization in Tanzania” Bank of Tanzania, Directorate of Economic Policy, May

Kihaule, M.K. (1994) “Behaviour of Demand for Money: a case of Tanzania” Unpublished dissertation, Department of Economics, University of Dar es Salaam, Tanzania

Maje, E.H. (1992) “Monetization, Financial Development and the Demand for Money: The Case of Tanzania” Unpublished dissertation, Department of Economics, University of Lund (Lund, Sweden)

McCallum, Bennett (1988), “Robustness Properties of a Monetary Rule”, Carnegie-Rochester Conference Series for Public Policy 29, Autumn: 173-203.

Mgonya, B. (1997) “Demand for Money and Infl ation in Tanzania 1966-1995” Unpublished dissertation, Department of Economics, University of Dar es Salaam, Tanzania

Mwase, Nkunde and Benno J. Ndulu (2007), “Tanzania” in B. J. Ndulu et al., eds, The Political Economy of Economic Growth in Africa, 1960-2000, Volume 1 (Cambridge: Cambridge University Press)

Ndanshau, M. (1996) “The Behaviour of Income Velocity in Tanzania, 1967-1994” African Economic Research Consortium, Nairobi

Ndulu, Benno J. (1997), “The Challenging Path of Transition to a Low Infl ation Economy in Tanzania” The Second Gilman Rutihinda Memorial Lecture (Dar es Salaam, Bank of Tanzania).

18


Nyella, Johnson J. (2005) “Seasonality of Currency in Circulation in Tanzania” Bank of Tanzania Technical Commentaries and Observations 08/2005 Dar es Salaam: Bank of Tanzania

Nyella, Johnson (2003), “Financial Programming: The Case of Tanzania” Dar es Salaam: Bank of Tanzania

Nyella, Johnson (1998) “The Demand for Money in Tanzania” Dar es Salaam: Bank of Tanzania

O’Connell, Stephen A. (1995), “Monetary Adjustment and Policy Compatibility in a Controlled Open Economy” Journal of African Economies 4(1): 52-82

Parastatal Sector Reform Commission (2000), The 1999/2000 Review and the Action Plan for 2000/2001, Government Printer, Dar es Salaam

Randa, John (1999), “Economic Reform and the Stability of the Demand for Money in Tanzania” Journal of African Economies 8(3): 307-344

Rashidi, Idris (1997), “Keynote Address to the 10th Conference of Financial Institutions: 7-9 April 1997, Arusha” in Bank of Tanzania Economic and Operations Report for the Year Ended 30th June 1997: 52-58

Siram, Subramanian (2001) “A Survey of Recent Empirical Money Demand Studies” IMF Staff Papers 47(3): 334-65

Waigama, Samuel M. S. (2008) Privatization Process and Asset Valuation: A Case Study of Tanzania Doctoral Dissertation, School of Architecture and the Built Environment, Royal Institute of Technology, Stockholm, Sweden

Walsh, Carl E. (2010) Monetary Theory and Policy, 3rd edition (Cambridge, MA: MIT Press)

19


Tables and FiguresTable 1. Principal components analysis, 1994q1-2008q4

Number of observations: 59, Number of components: 5

1. Principal components

Component Eigenvalue Difference Proportion Cumulative

Comp1 3.6546 2.8040 0.7309 0.7309

Comp2 0.8505 0.4151 0.1701 0.9010

Comp3 0.4354 0.4026 0.0871 0.9881

Comp4 0.0329 0.0063 0.0066 0.9947

Comp5 0.0266 . 0.0053 1.0000

2. Weights in fi rst principal component, and descriptive statistics

Variable Weights Mean Std. Dev. Min

gwssh 0.4233 0.0398 0.0060 0.0342

ytishr 0.4096 0.3778 0.0186 0.3480

dcpm3 0.4738 0.3541 0.1118 0.1482

gfcfgdp_di 0.5002 0.1927 0.0317 0.1492

impgdp_di 0.4224 0.2759 0.0629 0.2065

Notes: “Weights’ is the eigenvector that constitutes the fi rst principal component. See Appendix 3 for defi nitions and sources of variables.

Sources:STATA do fi le (structural_change.do) STATA data fi le (Tza_Md_qdata8908.dta)

20


Table 2. Long-run static regressions, M2 Dependent variable: l_m2r

Variable1994q1-2008q4 1998q1-2008q4

(1) (2) (3) (4)

l_rgne 0.6809 1.1111

7.53 7.49

l_rgdp 0.863 1.0169

12.14 8.22

spreada 0.0641 0.1194 -0.2734 0.0176

0.31 0.61 -1.01 0.08

l_infq -0.1543 0.2516 -0.783 -0.3383

-0.24 0.46 -1.67 -0.92

l_depfa -0.5252 -0.4707 -0.2415 -0.2713

-4.55 -4.14 -2.14 -3.23

struc_pc1 0.0957 0.0896 0.0607 0.0841

7.61 10.06 3.66 6.68

sq1 -0.0269 -0.0155 0.0015 0.0066

-1.29 -0.91 0.06 0.38

sq2 -0.0357 -0.0064 -0.0446 -0.0084

-1.48 -0.32 -1.99 -0.46

sq3 0.0018 0.0402 -0.0009 0.0274

0.06 1.90 -0.04 2.15

N 59 59 44 44

rmse 0.06199 0.05062 0.047 0.03727

aic -152.5 -176.4 -136.3 -156.7

bic -133.8 -157.7 -120.2 -140.6

ll 85.23 97.18 77.13 87.35

DF t-stat -3.961 -5.08 -4.198 -3.944

5% crit. val -3.43 -3.43 -3.43 -3.43

Notes: Robust t-statistics are below estimated coeffi cients. See Appendix 3 for defi nitions and sources of variables.

Sources: STATA do fi le (m2r_feb22_2010.do) STATA data fi le (Tza_Md_qdata8908.dta)

21


Table 3. Dynamic ECM regressions, M2 Dependent variable: D.l_m2r

Variable94q1-08q4 1998q1-2008q4

(1) (2) (3)

L.equi_err -0.3405 -0.3324 -0.3294

-2.88 -2.52 -3.28

LD.l_m2r 0.3398 0.3103 0.3039

2.82 1.58 1.70

D.l_rgdp 0.126 0.4317 0.4355

1.03 2.29 3.00

LD.l_rgdp -0.1616 -0.0084

-1.33 -0.08

D.spreada 0.0631 -0.0719

0.64 -0.27

LD.spreada -0.0047 0.1086

-0.04 0.51

D2.spreada -0.0862

-0.58

D.l_infq -0.3235 -0.5002 -0.5087

-1.2 -1.84 -2.71

LD.l_infq 0.0818 0.0191

0.37 0.10

D.l_depfa -0.0533 -0.0862 -0.0847

-0.83 -1.66 -1.87

LD.l_depfa 0.1836 0.0084

1.93 0.06

sq1 -0.0322 -0.0145 -0.0147

-2.22 -0.86 -1.06

sq2 0.018 0.0102 0.0095

1.32 0.46 0.50

sq3 0.0322 0.0432 0.0424

2.12 2.4 2.62

N 58 42 42

Rmse 0.02895 0.02525 0.02364

Aic -234.3 -178.9 -186.8

Bic -205.5 -154.5 -169.4

Ll 131.2 103.4 103.4

P-values

Res normality 0.05 0.001 0.001

BP Het 0.038 0.008 0.009

BG Auto lag1 0.45 0.878 0.997

lag2 0.016 0.281 0.367

lag3 0.014 0.106 0.166

lag4 0.009 0.181 0.271

Note: l.equi_err is the lagged equilibrium error from the static regression in column 2 or 4 of Table 2.

Sources: see Table 2.

22


Table 4. Long-run static regressions, Currency and DepositsDependent variable: l_cur or _depr

Currency Deposits

Variable 94q1-08q4 98q1-08q4 94q1-08q4 98q1-08q4

(1) (2) (3) (4)

l_rgdp 0.5247 0.5636 0.9734 1.2427

7.03 3.57 13.15 9.45

spreada -0.0789 -0.3198

-0.43 -1.23

l_infq -0.3122 -0.9283 0.4406 -0.1385

-0.47 -2.39 0.77 -0.32

l_depfa -0.4778 -0.096 -0.4193 -0.3028

-3.15 -0.76 -3.47 -2.94

struc_pc1 0.0831 0.0935 0.1008 0.0822

9.46 5.81 11.25 6.44

sq1 -0.0893 -0.0688 0.0192 0.043

-4.36 -3.29 0.95 1.98

sq2 -0.0714 -0.0606 0.0184 0.0137

-2.79 -2.67 0.88 0.66

sq3 -0.0063 -0.0124 0.0594 0.0465

-0.23 -0.54 2.64 2.84

N 59 44 59 44

rmse 0.06168 0.04703 0.05536 0.04333

aic -153.9 -137 -165.8 -143.4

bic -137.3 -122.7 -147.1 -127.4

ll 84.94 76.49 91.9 80.71

DF t-stat -4.049 -3.995 -5.245 -4.274

5% crit. val -3.43 -3.43 -3.43 -3.43

Notes: Robust t-statistics are below estimated coeffi cients. See Appendix 3 for defi nitions and sources of variables.

Sources:STATA do fi le (m2r_feb22_2010cur_rv.do and m2r_feb22_2010cur_rv.do) STATA data fi le (Tza_Md_qdata8908.dta)

23


Table 5. Dynamic ECM regressions, CurrencyDependent variable: D.l_cur

Variable94q1-08q4 1998q1-2008q4

(1) (2) (3)

L.equi_err -0.3942 -0.4971 -0.5196

-3.31 -2.71 -3.12

LD.l_cur 0.1872 0.2184 0.2192

1.37 1.11 1.43

D.l_rgdp 0.2399 0.4336 0.4432

2.06 1.71 1.76

LD.l_rgdp -0.1359 0.0676

-1.31 0.35

D.l_ifwd2a -2.1789 -7.5882 -5.5459

-1.87 -1.78 -1.82

LD.l_ifwd2a 2.0943 3.1821

1.48 0.77

D.l_infq -0.6995 -0.97 -0.8423

-2.06 -2.46 -2.88

LD.l_infq -0.1824 -0.198

-0.57 -0.56

D.l_depfa -0.1246 -0.1309

-1.44 -1.14

LD.l_depfa 0.3162 0.2672

2.00 1.07

D2.l_depfa -0.1956

-1.59

sq1 -0.1029 -0.0856 -0.0848

-5.92 -2.88 -3.07

sq2 0.0153 0.0202 0.0267

0.48 0.57 0.77

sq3 0.0403 0.026 0.0367

2.13 1.00 1.78

N 58 42 42

Rmse 0.03878 0.0409 0.03898

Aic -200.4 -138.3 -144.8

Bic -171.6 -114 -127.4

Ll 114.2 83.17 82.39

P-values

Res normality 0.576 0.516 0.647

BP Het 0.811 0.955 0.766

BG Auto lag1 0.096 0.131 0.25

lag2 0.233 0.071 0.06

lag3 0.391 0.135 0.124

lag4 0.091 0.196 0.201


Source: see Table 4.

24


Table 6. Dynamic ECM regressions, DepositsDependent variable: D.l_depr

Variable94q1-08q4 1998q1-2008q4

(1) (2) (3)

L.equi_err -0.3343 -0.2714 -0.3196

-2.59 -2.20 -3.55

LD.l_depsr 0.4298 0.2856 0.3074

3.86 1.98 2.31

D.l_rgdp 0.0948 0.3906 0.4168

0.70 2.21 3.26

LD.l_rgdp -0.1463 0.0235

-1.09 0.15

D.spreada -0.0391 -0.2947 -0.2747

-0.34 -1.02 -1.32

LD.spreada 0.0695 0.2656

0.56 1.69

D.l_infq -0.2924 -0.4104 -0.4302

-1.13 -1.50 -2.07

LD.l_infq 0.0192 0.0073

0.09 0.04

D.l_depfa -0.0322 -0.0713 -0.0718

-0.41 -1.12 -1.14

LD.l_depfa 0.1338 -0.0892

1.08 -0.53

sq1 0.0092 0.0158 0.019

0.61 1.11 1.49

sq2 0.0067 -0.0057 -0.0099

0.53 -0.33 -0.79

sq3 0.0308 0.0417 0.0382

1.95 2.49 2.62

N 58 42 42

Rmse 0.03289 0.02746 0.0263

Aic -219.5 -171.8 -177.8

Bic -190.7 -147.5 -160.5

Ll 123.8 99.91 98.92

P-values

Res normality 0.027 0.001 0.001

BP Het 0.765 0.977 0.929

BG Auto lag1 0.817 0.19 0.768

lag2 0.028 0.222 0.142

lag3 0.016 0.084 0.083

lag4 0.009 0.109 0.139


Source: see Table 4.

25


Table 7. Velocity: within-sample predictive performanceSample 2002q1 to 2008q4

Model (dependent variable log velocity)

RMSE( % per quarter)

MAE(% per quarter)

Rolling Trend 5.26 4.47

Seasonal Moving Average 6.66 5.16

Random walk with seasonals 4.09 3.13

SARIMA (1,1,2,4) 3.61 2.92

VAR (2) 2.42 1.70

Notes: Measures indicate the accuracy of the within-sample predictive power of the model, where for each model the prediction error is ]. RMSE is the root mean square error; MAE the mean absolute error. SARIMA denotes seasonal autoregressive integrated moving average of order p=1, d=1, q=2 and s=4 where p denotes the autoregressive order, d the order of integration, q the moving average error order and s the number of deterministic seasonal dummy variables. VAR(2) denotes vector auto-regression with two lags on each variable (plus deterministic seasonal dummy variables).

Table 8. Velocity: out-of-sample forecast performanceEstimation Sample 1998q1 to 2006q4Forecast Sample 2007q1 to 2008q4

Model (dependent variable log velocity)

1-step-ahead forecast (% per quarter)

8-step-ahead forecast (% per quarter)

RMSE MAE RMSE MAE

Rolling Trend 6.04 5.61 - -

Seasonal Moving Average 7.67 6.99 - -

Random walk w/ seasonals 3.76 3.20 7.45 6.95

SARIMA (1,1,2,4) 4.05 3.05 7.31 6.86

VAR (2) 2.51 2.11 5.13 4.54

Notes: See Table 7 and text.

26


Figure 1. Velocity of monetary aggregates, 1988q1-2008q4

Source: Bank of Tanzania

Figure 2. Composition of M3


120.0

100.0

80.0

60.0

40.0

20.0

01981 1986 1991 1996 2001 2006

cu m1

m2 m3

4 4.

5 5

5.5

6 6.

5

0 10

20

30

40

50

1990q1 1995q1 2000q1 2005q1 2010q1 1990q1 1995q1 2000q1 2005q1 2010q1

1990q1 1995q1 2000q1 2005q1 2010q1 1990q1 1995q1 2000q1 2005q1 2010q12

2.5

3 3.

5

0 20

40

60

1.2

1.4

1.6

1.8

2 2.

2

0 20

40

60

.8

1 1.

2 1.

4 1.

6 1.

8

10

20

30

40

50

Blue: Velocity (Left) Red: Y/Y Money Growth (Right)

27


Figure 3. Currency and food price seasonals

Pf/Pn C/Pn Food price inflation (t+2)

Source: NBS and Bank of Tanzania. Data are from 2001m1 to 2008m12. Pf and Pn are the food and non-food consumer price indexes, and C is currency held outside the banking system. Seasonal factors are multiplicative for C/Pf and Pf/Pn and additive moving-average for food price infl ation.

Figure 4. Ex post yields on domestic and foreign currency deposits

Expected return on FCD DC_12m_td rate


106.0

104.0

102.0

100.0

98.0

96.0

94.0

92.0

90.0

88.0

2.50

2.00

1.50

1.00

0.50

0.00

-0.50

-1.00

-1.50

-2.00

-2.50Jan Feb Mar April May June July Aug Sept Oct Nov Dec

25

20

15

10

5

0

-5

-10

%

June

99

Nov

99

Ap

ril 0

0

Sep

t 00

Feb

01

July

01

Dec

01

May

02

Oct

02

Mar

03

Aug

03

Jan

04

June

04

Nov

04

Ap

ril 0

5

Sep

t 05

Feb

06

July

06

Dec

06

May

07

Oct

07

Mar

ch 0

8

Aug

08

28


Figure 5. Structural Change and the Monetary Intensity of Economic Activity.

Note: See Table 1 for details on the principal components calculation.

Figure 6. Long-run equilibrium: actual and fi tted values

Tanzania log real Money M2 1998q1 to 2008q4 Actual vs Fitted

+/- 2s.e. fitted log real M2

Source: Column 4 of Table 2.

.34

.36

.38

.4

.42

Modem share of GDP

1995q1 2000q1 2005q1 2010q1

Investment (% GDP)

1995q1 2000q1 2005q1 2010q1

Imports (% GDP)

1995q1 2000q1 2005q1 2010q1

.15

.2

.25

.2

.25

.3

.35

.4

Domestic credit to private sector (%M

1995q1 2000q1 2005q1 2010q1

.1

.2

.3

.4

.5

.6

Govt wage bill (% GDP)

1995q1 2000q1 2005q1 2010q1

.035

.04

.0

45

.05

.055

-2

0 2

4

1995q1 2000q1 2005q1 2010q1

1998q1 2000q1 2002q1 2004q1 2006q1 2008q1

9 9.

5 10

10

.5lo

g re

al M

2

29


Figure 7. Short-run dynamics: actual and fi tted values

Tanzania log real Money M2 1998q1 to 2008q4 Actual vs Fitted

+/- 2s.e. fitted growth in log real M2

Source: Column 3 of Table 3.

Figure 8. Time-series forecasts of M2 velocity

Log M2 velocity 1998q1 to 2008q4

log velocity [LHS] Quarter on Quarter change in velocity [RHS]

Source: Bank of Tanzania.

-.2

-.1

0 .1

.2

1998q1 2000q1 2002q1 2004q1 2006q1 2008q1

Scale variable: rgdp

0 20

40

60

80

1998q1 2000q1 2002q1 2004q1 2006q1 2008q1

Gro

wth

in lo

g re

al M

2 (%

)lo

g r

eal M

2

80 60

40 20

0G

row

th in log

real M2 (%

)

30


Figure 9. Velocity forecasting models: within-sample performance

(a) Rolling trend forecaster vs random walk – Log M2 velocity 2004q1 to 2008q4Within sample prediction error

Rolling trend Random Walk

(b) ARIMA versus VAR – Log M2 velocity 2004q1 to 2008q4Within sample prediction error

ARIMA VAR

2004q1 2005q1 2006q1 2007q1 2008q1 2009q1

-10

-5

0 5

10

2004q1 2005 q1 2006q1 2007q1 2008q1 2009q1

-10

-5

0 5

10lo

g re

al M

2lo

g re

al M

2

31


Figure 10. One-step-ahead forecast error from alternative forecast models

Log M2 velocity out-of-sample forecast performanceOne-step-ahead forecast errors

Rolling trend SARIMA VAR

Figure 11. Actual velocity versus 8-step ahead VAR-based forecast

Log M2 velocity: 2007q1 to 2008q4VAR-based dynamic forecast

95% CI forecast observed

-10

-5

0 5

10

2006q1 2006q3 2007q1 2007q3 2008q1 2008q3

-10

-5

0 5

10

2006q3 2007q1 2007q3 2008q1 2008q3

log

real

M2

log

real

M2

32


Appendix 1. Macroeconomic Developments, 1966-95The Tanzanian shilling was introduced in 1966 in the context of a fi xed exchange rate regime. With the Arusha Declaration in 1967, Tanzania embarked on a socialist path that gave the National Bank of Commerce, a state-owned enterprise, a near-monopoly over the commercial banking sector and produced rapid growth in other parts of the public sector. In a context of emerging macroeconomic imbalances and infl ationary pressures, the fi rst oil shock (1973/74) produced a balance of payments crisis. Reluctant to devalue the exchange rate, the BOT responded by tightening exchange controls, which led to the emergence of a substantial parallel market for foreign exchange (Figure A1). As infl ationary pressures continued during the 1970s the government responded with administrative measures, imposing a broadening array of price controls on widely-consumed items.

Figure A1. Infl ation, depreciation, and parallel premium (%)

Inflation Official rate Premium

The 1976/77 coffee boom provided temporary relief from macroeconomic imbalances, but at the end of the 1970s a set of macroeconomic shocks – the end of the coffee boom, the onset of the second global oil crisis (1978/79), and the fi nancing of a war with Uganda – produced a loss of fi scal control and a renewed balance of payments crisis. Fiscal imbalances were monetized – broad money grew by 38 percent in 1979 alone – and excess demand pressures led to a further tightening of controls on foreign exchange allocation and domestic prices. The combination proved unsustainable, more so following the global debt crisis of 1982 and a souring of relations between the government of Tanzania and its major donors. Economic growth virtually stopped in the late 1970s, averaging 0.8 percent between 1977 and 1983 and turning sharply negative in per-capita terms, amid a general deterioration of social services.12

A maxi-devaluation in 1983, undertaken in the midst of deep macroeconomic diffi culties, marked the beginning of what ultimately proved a decisive transition from direct controls to price-based mechanisms of macroeconomic adjustment in Tanzania. President Nyerere resigned in 1985, and in the following year the country embarked on a sequence of 3-year Economic Recovery Programmes (ERP I, 1986-89, and ERP II, 1989-92) with IMF and World Bank support. These were designed to convert an administratively controlled economy into one with market-determined prices and private ownership and control. Over the course of a decade, these reforms transformed the banking sector and the menu of fi nancial assets available to Tanzanian households.

100.0

80.0

60.0

40.0

20.0

0

-20.01981 1986 1991 1996 2001 2006

350.0

300.0

250.0

200.0

150.0

100.0

50.0

0.0

12 The offi cial economy shrank rapidly during the early 1980s as economic agents responded to scarcity by diverting dwindling supplies of goods and foreign exchange onto parallel markets. The macroeconomics of the control period 1967-1985 have been studied extensively, including the implosion of the offi cial economy after 1980 due to shortages, rationing, and a monetary overhang. See Bevan et al. 1993 and O’Connell 1997.

33


Financial Sector Reforms13

Financial sector reforms began in 1987 with interest-rate adjustments designed to move real interest rates into positive territory (Figure 1). The banking system nonetheless remained heavily repressed until the early 1990s. All fi nancial institutions were fully owned by the government, and entry of private and foreign banks was prohibited. As late as 1993, there were only 3 banks in operation, one of which operated in Zanzibar; the National Bank of Commerce accounted for over 90 percent of deposits.14 The government controlled interest rates and the exchange rate, and and allocated credit and foreign exchange administratively. The burden of non-performing loans reached 60 percent in 1992, refl ecting poor accountability and an absence of prudential norms.

Figure A2. Infl ation and interest rates

Inflation T-bill rate Savings deposits

Time deposits

Following the Nyirabu Commission’s report in 1990, the government undertook a sequence of reforms that transformed an insolvent nationalized monopoly banking system into an increasingly competitive sector dominated by foreign-owned banks (Cull and Spreng 2008). The process began with the entry of the fi rst private foreign banks in 1992 and 1993 (they began operations in 1994) and the privatization of the Cooperative and Rural Development Bank (CRDB) in 1996. The dominant National Bank of Commerce (NBC) was recapitalized in successive efforts during the early and mid-1990s, and in 1997 the Government divided it into the new NBC, which inherited the bank’s urban branches and its traditional urban- and trade-based lending business, and the National Microfi nance Bank (NMB) which inherited the old NBC’s network of rural branches and its mandate to provide banking services to rural populations. The new NBC was privatized in 2000, and the NMB went 0 to 49% private participation in 2005.

13 See Nyella (2003)14 National Bank of Commerce (NBC), Cooperative and Rural Development Bank (CRDB) and the People’s Bank of Zanzibar (PBZ). In

July 1997, the NBC was split into two banks, the NBC and the National Microfi nance Bank (NMB).

1981 1986 1991 1996 2001 2006

50

40

30

20

10

0

34


The introduction of competition transformed the quality, price, and availability of banking services in Tanzania, with potentially large impacts on the demand for bank liabilities. The evidence suggests substantial increases in bank profi tability and effi ciency. The restructuring process also appears to have restricted the access of households to fi nancial services, however, at least initially and particularly in rural areas; the NBC retrenched 2,800 employees and closed 23 branches in 1994 alone. Successive household budget surveys show a sharp decrease during the 1990s in the proportion of households reporting a bank account – this number falls from 18.0 percent in the 1991/92 survey to 6.4 percent in 2000/01, before recovering to 10.0 percent in the 2007 survey.15 But the Banking and Financial Institutions Act of 1991 authorized the entry of private and foreign banks, and in 1994 the fi rst two (Standard Chartered and Meridien–BIAO, both foreign-owned) began operations. By the end of 2001, the total number of banks had reached 17; this number doubled again by 2009, with the majority of Tanzania’s 34 banks foreign-owned., along with a possible reduction in access to fi nancial services in rural areas, for example through the closure of branches when NBC was split (Cihak and Podpiera 2005). The rapid introduction of mobile banking and other technological innovations over the past 5 years or so has further transformed the menu of banking services available in Tanzania.

A set of contemporaneous reforms during the early 1990s converted a system of exchange controls and multiple exchange rates into one in which private residents were allowed to hold foreign currency and domestic foreign-currency deposits and exchange rates were determined in an inter-bank market. The Foreign Exchange Act of 1992 introduced privately-owned foreign exchange shops and allowed residents to hold foreign currency deposits for the fi rst time. A weekly foreign exchange auction system was introduced by the BOT in 1993, with the offi cial exchange rate determined by the marginal market clearing bid; as intended, this was replaced in June 1994 by the Inter-bank Foreign Exchange Market (IFEM).

Monetary and fi scal developments

The early 1990s saw a gradual shift from direct to indirect instruments of monetary control, facilitated by interest rate liberalization and the introduction of Treasury bill auctions in 1993. The ERPs emphasized macroeconomic stabilization and introduced annual money-growth targets as a means of infl ation control. Infl ation remained persistently high, however, as fi scal pressures drove money growth rates well above program targets (Figure 1). Thus M2 growth averaged 32.4 percent during the reform period, as against the target range of 10 to 17 percent, the difference largely refl ecting government borrowing to fi nance crop authorities and cooperative unions (Ndanshau 1996).

Monetary accommodation continued over the fi ve-year period following the 1990 general elections. In a context of expansionary fi scal policy, the growth rate of extended broad money averaged 36.3 percent and infl ation averaged nearly 30 percent (Figure 2). By 1994/95, sharply rising interest costs had convinced the policy authorities that fi scal adjustment was essential for macroeconomic stabilization (Rashidi, 1997).

15 It is likely that the proportion bottomed out in the mid- to late-1990s, as NBC was being restructured and new banks had not yet established a substantial foothold. Reductions in rural access may well have lasted longer, however. The National Microfi nance Bank, which succeeded NBC and acquired most of NBC’s rural branch network, remained in state hands until 2005, operating for most of the 1997-2005 period under a private consulting contract that focused on enhancing profi tability, not on expansion of reach or services. See Cull and Spreng 2008.

35


Figure A3. Infl ation and money growth

Inflation M2 M3

The conduct of monetary and fi scal policy changed abruptly in 1995. The Bank of Tanzania Act placed price stability fi rst among the objectives of monetary policy and granted the BOT the independence required to carry out its role and functions. The government adopted a cash budget, accepting a virtually complete denial of monetary fi nance. As the cash budget and reserve money programs gained traction, the growth rate of extended broad money fell sharply, averaging 13.2 percent for the fi ve year period ending 2000. Disinfl ation was rapid; infl ation fell from 30% in 1995 to 16% in 1997 and below 10% in 1999 (Figure A3).

1981 1986 1991 1996 2001 2006

50

40

30

20

10

0

36


Appendix 2. Motivating the demand for moneyThe demand for any monetary aggregate can be motivated by reference to the portfolio problem of the household sector (e.g., Walsh 2010).

The end-of-period fi nancial wealth of the private non-bank sector is given by

(A2.1)

where C and D refer to cash and bank deposits, E is the exchange rate in Tanzanian shillings per US dollar, $ denotes holdings in foreign currency, and B denotes bonds, which on the domestic side are restricted to government securities and on the foreign side are illegal for Tanzanian citizens. More generally, the private sector’s overall wealth W also includes real assets, including inventories of goods.

Measuring wealth in real terms as w, the private sector’s budget constraint in terms of domestic goods can be written

(A2.2)

where lower-case letters refer to quantities measured in real terms and r j denotes a real return to asset j. The opportunity cost of holding wealth as domestic money varies across the components of money and the alternative assets that can be held. For domestic asset m and alternative asset x, therefore,

Opportunity cost of m in terms of x =

(A2.3)

The opportunity cost of currency or non-interest-bearing bank deposits in terms of domestic bonds, for example, is where is the infl ation rate; as long as is small, this is approximately the nominal interest rate .The opportunity cost of domestic currency in terms of foreign currency is ; if

is small this is approximately , the sum of the foreign interest rate and the rate of exchange rate depreciation. Interest-bearing deposits pay an ‘own return’, and this reduces their opportunity cost in terms of other assets; the opportunity cost of time deposits in terms of government bonds, for example, is approximately .

Note that although nominal interest rates are generally known at the time portfolio decisions are made, the only real returns known in advance are those on assets that are explicitly indexed to domestic infl ation. Assets denominated in foreign currency are subject to exchange risk, and interest-bearing assets are subject to infl ation risk (and, for long-term bonds, risk of capital gain or loss due to interest rate changes). Narrowly-defi ned real assets are subject to capital gains or losses when relative prices change, but the average capital gain on a diversifi ed basket of goods is zero. An inventory of goods can therefore serve as an effective hedge against infl ation if storage costs are low.

37


Except for currency, the monetary aggregates include a mixture of assets with different own returns. There is therefore no single opportunity cost of M2 in terms of any other asset. Since the opportunity costs can be approximately reduced to linear combinations of known or expected nominal yields, the standard empirical approach is to allow the demand for each aggregate to depend on the full set of expected nominal yields on own and alternative assets.

Since opportunity costs are typically positive, some other features of domestic money must motivate its being held. Primary among these is the convenience of using money for transactions. The demand for money therefore also depends on the anticipated volume of transactions, which we proxy using real GDP or real expenditure. Real wealth is an important additional variable in most treatments of money demand, but it is not observable in Tanzania and so we omit it.

38


Appendix 3. Defi nitions and sources of variablesVariable name Defi nition Details and source

l_cur log(Currency Outside Banks/CPI)

End-of-quarter Currency Ouside Banks (BOT); Quarterly average CPI (NBS)

l_depr log(Domestic Currency Deposits/CPI)

End-of-quarter Domestic Currency Deposits (BOT)

l_m2r log(M2/CPI) End-of-quarter M2 (BOT)

l_rgdp log(Real GDP) NBS and interpolated to quarterly frequency by authors.

l_rgne log(Real Gross National Expenditure)

NBS and interpolated to quarterly frequency by authors.

l_ifwd2a log(1 + iWD /100) iWD is the weighted interest rate on domestic-currency deposits (weights correspond to weights of saving and time deposits in total deposits). Quarterly average interest rates (BOT)

l_infq log(1+π100) π is the CPI infl ation rate between quarters t and t-1

l_depfa log(1+x/100) x is the rate of depreciation for average end of quarter exchange rate TZS/ USD (BOT)

spreada log(1 + iTB/100) − ln(1 + iWD/100)

iTB is the 3-month treasury bill interest rate (BOT)

struc_pc1 First principal component of the following 5 transactions intensity indicators:

gwssh Government wages and salaries as a share of GDP (BOT)

ytishr Share of mining and quarrying; manufacturing; electricity and water supply; trade, restaurants and hotels; transport and communications; and fi nance, insurance and real estate, in GDP (NBS).

dcpm3 Domestic credit to the private sector as a share of M3 (BOT)

gfcfgdp_di Gross fi xed capital formation as a share of GDP (NBS)

impgdp_di Imports as a share of GDP (NBS)

39


Appendix 4. Phillips-Perron Unit root tests b_tr t_tr p(z)_tr p(z)_dr

l_m2r 0.002 2.733 0.912 1.000

D.l_m2r 0.001 3.555 0.000 0.000

l_cur 0.003 2.259 0.904 0.998

D.l_cur 0.001 1.193 0.000 0.000

l_depsr 0.002 3.992 0.622 1.000

D.l_depsr 0.001 4.754 0.000 0.000

l_rgne 0.010 5.028 0.000 0.995

D.l_rgne 0.001 1.740 0.000 0.000

l_rgdp 0.012 6.282 0.000 0.996

D.l_rgdp 0.001 1.350 0.000 0.000

l_ifsda 0.000 0.232 0.918 0.205

D.l_ifsda 0.000 2.026 0.000 0.000

l_iftda 0.000 0.358 0.876 0.408

D.l_iftda 0.000 1.384 0.000 0.000

l_infa 0.000 0.089 0.956 0.564

D.l_infa 0.000 1.731 0.000 0.000

l_depfa -0.000 -0.791 0.007 0.001

D.l_depfa 0.000 1.165 0.000 0.000

l_e 0.004 2.420 0.351 0.611

D.l_e -0.000 -0.737 0.000 0.000

l_infq -0.000 -2.823 0.000 0.001

D.l_infq 0.000 0.869 0.000 0.000

Spreada -0.000 -0.093 0.021 0.002

D.spreada 0.000 0.214 0.000 0.000

struc_pc1 0.014 11.817 0.018 0.993

D.struc_pc1 0.005 3.302 0.008 0.477

Notes: The columns of the table show the results of Phillips-Perrron unit root tests applied to variables on the rows (the null hypothesis is a unit root). The fi rst 3 columns refer to PP tests with a trend term included: b_tr and t_tr are the estimated coeffi cient on the trend term and its t-statistic, and p(z)_tr is the MacKinnon approximate p value for the z(t) statistic. The fi nal column is the MacKinnon approximate p value for tests with a constant but no trend term. For any chosen signifi cance value , values of p(z) less than imply rejection of a unit root at the signifi cance level .

Key: l_x and D.l_x refer to the level and fi rst difference of the log of variable x.

Directed and Organised by

Initiated and Funded by

About the International Growth CentreThe IGC offers independent advice on economic growth to governments of developing countries. Based at the London School of Economics and in partnership with Oxford University, the IGC is initiated and funded by the UK Department for International Development (DFID).

The IGC has active country programmes in Bangladesh, Ethiopia, Ghana, India, Pakistan, Sierra Leone, Tanzania, Mozambique, Zambia and Rwanda and supports over seventy individual research projects on issues of governance, human capital, agriculture, infrastructure, trade, firm capability, state capacity, macroeconomics and political economy.

The IGC is directed by a Steering Group consisting of an Executive Director (Gobind Nankani) in collaboration with a Deputy Executive Director (Mark Henstridge) and two Academic Directors, one from LSE (Robin Burgess) and one from Oxford University (Paul Collier). The Steering Group also includes Chang-Tai Hsieh from the University of Chicago, Timothy Besley at LSE and Stefan Dercon at Oxford University.

The organisational structure of the IGC spans a London hub, country offi ces in partner countries, a group of 10 research programmes with participation from academics in world-class institutions, a network of policy stakeholders in the developing world and a range of public, civil society and private sector partners.

Contact usInternational Growth CentreThe London School of Economics and Political Science4th Floor, Tower TwoHoughton StreetLondon WC2A 2AEUnited Kingdom

General Offi ce Tel: +44 (0)20 7955 6144

For enquiries about this paper, please contact Adam Green:[email protected] +44 (0)20 7955 3665

Documents

The Demand for Money in Tanzania · both serve as proxies for the transactions demand for money.2 The two variables behave similarly over time, but differences due to aid, terms of